[Bug rtl-optimization/68955] [6 Regression] wrong code at -O3 on x86-64-linux-gnu in 32-bit mode

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68955

--- Comment #14 from Jakub Jelinek  ---
Author: jakub
Date: Thu Feb 11 09:23:06 2016
New Revision: 21

URL: https://gcc.gnu.org/viewcvs?rev=21=gcc=rev
Log:
Backported from mainline
2016-01-19  Jakub Jelinek  

PR rtl-optimization/68955
PR rtl-optimization/64557
* dse.c (record_store, check_mem_read_rtx): Don't call get_addr
here.  Fix up formatting.
* alias.c (get_addr): Handle VALUE +/- CONST_SCALAR_INT_P.

* gcc.dg/torture/pr68955.c: New test.

Added:
branches/gcc-4_9-branch/gcc/testsuite/gcc.dg/torture/pr68955.c
Modified:
branches/gcc-4_9-branch/gcc/ChangeLog
branches/gcc-4_9-branch/gcc/alias.c
branches/gcc-4_9-branch/gcc/dse.c
branches/gcc-4_9-branch/gcc/testsuite/ChangeLog

[Bug rtl-optimization/68955] [6 Regression] wrong code at -O3 on x86-64-linux-gnu in 32-bit mode

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68955

--- Comment #13 from Jakub Jelinek  ---
Author: jakub
Date: Wed Feb 10 18:34:30 2016
New Revision: 233293

URL: https://gcc.gnu.org/viewcvs?rev=233293=gcc=rev
Log:
Backported from mainline
2016-01-19  Jakub Jelinek  

PR rtl-optimization/68955
PR rtl-optimization/64557
* dse.c (record_store, check_mem_read_rtx): Don't call get_addr
here.  Fix up formatting.
* alias.c (get_addr): Handle VALUE +/- CONST_SCALAR_INT_P.

* gcc.dg/torture/pr68955.c: New test.

Added:
branches/gcc-5-branch/gcc/testsuite/gcc.dg/torture/pr68955.c
Modified:
branches/gcc-5-branch/gcc/ChangeLog
branches/gcc-5-branch/gcc/alias.c
branches/gcc-5-branch/gcc/dse.c
branches/gcc-5-branch/gcc/testsuite/ChangeLog

[Bug rtl-optimization/68955] [6 Regression] wrong code at -O3 on x86-64-linux-gnu in 32-bit mode

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68955

Jakub Jelinek  changed:

   What|Removed |Added

 Status|ASSIGNED|RESOLVED
 Resolution|--- |FIXED

--- Comment #12 from Jakub Jelinek  ---
Fixed.

[Bug rtl-optimization/68955] [6 Regression] wrong code at -O3 on x86-64-linux-gnu in 32-bit mode

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68955

--- Comment #11 from Jakub Jelinek  ---
Author: jakub
Date: Tue Jan 19 12:34:45 2016
New Revision: 232554

URL: https://gcc.gnu.org/viewcvs?rev=232554=gcc=rev
Log:
PR rtl-optimization/68955
PR rtl-optimization/64557
* dse.c (record_store, check_mem_read_rtx): Don't call get_addr
here.  Fix up formatting.
* alias.c (get_addr): Handle VALUE +/- CONST_SCALAR_INT_P.

* gcc.dg/torture/pr68955.c: New test.

Added:
trunk/gcc/testsuite/gcc.dg/torture/pr68955.c
Modified:
trunk/gcc/ChangeLog
trunk/gcc/alias.c
trunk/gcc/dse.c
trunk/gcc/testsuite/ChangeLog

[Bug rtl-optimization/68955] [6 Regression] wrong code at -O3 on x86-64-linux-gnu in 32-bit mode

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68955

Jakub Jelinek  changed:

   What|Removed |Added

  Attachment #37362|0   |1
is obsolete||

--- Comment #10 from Jakub Jelinek  ---
Created attachment 37385
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=37385=edit
gcc6-pr68955.patch

Untested updated patch.

[Bug rtl-optimization/68955] [6 Regression] wrong code at -O3 on x86-64-linux-gnu in 32-bit mode

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68955

--- Comment #8 from Jakub Jelinek  ---
So, we have before DSE2:
...
(insn 525 192 458 12 (set (reg:SI 1 dx [orig:247 ivtmp.44 ] [247])
(mem/c:SI (plus:SI (reg/f:SI 6 bp)
(const_int -76 [0xffb4])) [4 %sfp+-52 S4 A32]))
pr68955.c:27 86 {*movsi_internal}
 (nil))
(insn 458 525 524 12 (set (reg:SI 0 ax [orig:247 ivtmp.44 ] [247])
(reg:SI 1 dx [orig:247 ivtmp.44 ] [247])) pr68955.c:27 86
{*movsi_internal}
 (nil))
(insn 524 458 194 12 (set (reg:SI 4 si [orig:124 _95 ] [124])
(mem/c:SI (plus:SI (reg/f:SI 6 bp)
(const_int -80 [0xffb0])) [4 %sfp+-56 S4 A32]))
pr68955.c:27 86 {*movsi_internal}
 (nil))
(insn 194 524 523 12 (set (mem:SI (reg:SI 0 ax [orig:247 ivtmp.44 ] [247]) [2
MEM[base: _165, offset: 0B]+0 S4 A32])
(reg:SI 4 si [orig:124 _95 ] [124])) pr68955.c:27 86 {*movsi_internal}
 (nil))
(note 523 194 567 12 NOTE_INSN_DELETED)
(insn 567 523 460 12 (set (reg:SI 2 cx [orig:125 pretmp_96 ] [125])
(mem/c:SI (plus:SI (reg/f:SI 6 bp)
(const_int -84 [0xffac])) [4 %sfp+-60 S4 A32]))
pr68955.c:28 86 {*movsi_internal}
 (nil))
(insn 460 567 196 12 (set (reg:SI 0 ax [orig:125 pretmp_96 ] [125])
(reg:SI 2 cx [orig:125 pretmp_96 ] [125])) pr68955.c:28 86
{*movsi_internal}
 (nil))
(insn 196 460 598 12 (set (mem/c:SI (const:SI (plus:SI (symbol_ref:SI ("i")
[flags 0x2]  )
(const_int 308 [0x134]))) [2 i+308 S4 A32])
(reg:SI 0 ax [orig:125 pretmp_96 ] [125])) pr68955.c:28 86
{*movsi_internal}
 (nil))
(insn 598 196 197 12 (set (reg:SI 0 ax [orig:209 ivtmp.42 ] [209])
(mem/c:SI (plus:SI (reg/f:SI 6 bp)
(const_int -28 [0xffe4])) [4 %sfp+-4 S4 A32]))
pr68955.c:26 86 {*movsi_internal}
 (nil))
(insn 197 598 198 12 (set (reg:CCGC 17 flags)
(compare:CCGC (reg:SI 5 di [orig:122 _93 ] [122])
(mem:SI (plus:SI (reg:SI 0 ax [orig:209 ivtmp.42 ] [209])
(const_int 20 [0x14])) [2 MEM[base: _327, offset: 20B]+0 S4
A32]))) pr68955.c:26 7 {*cmpsi_1}
 (nil))
...
and the bug IMHO is that the read in 197 can alias the store in insn 194 (well,
in the testcase for e == 0, h == 2 it actually is some later read, insn 194 is
from unrolled k == 0, 197 from k == 1, while the actual problem on the testcase
occurs on k <= 4 store vs. k == 5 first read, but the alias oracle really
doesn't know this), but during
check_mem_read_rtx we call canon_true_dependence with:
(gdb) p debug_rtx (mem)
(mem:SI (reg:SI 0 ax [orig:247 ivtmp.44 ] [247]) [2 MEM[base: _165, offset:
0B]+0 S4 A32])
(gdb) p mem_mode
$41 = SImode
(gdb) p debug_rtx (mem_addr)
(reg:SI 0 ax [orig:247 ivtmp.44 ] [247])
(gdb) p debug_rtx (x)
(mem:SI (plus:SI (reg:SI 0 ax [orig:209 ivtmp.42 ] [209])
(const_int 20 [0x14])) [2 MEM[base: _327, offset: 20B]+0 S4 A32])
(gdb) p debug_rtx (x_addr)
(plus:SI (reg:SI 0 ax [orig:209 ivtmp.42 ] [209])
(const_int 20 [0x14]))
arguments and that tells us that the two can't alias, because
memrefs_conflict_p returns that they don't.  That is because it sees register
%eax used as base address of both, one memory is biased by offset 20 and the
other is not, and both sizes are 4 bytes.  That is true, except the register
contains different value in between the two instructions.
So, the question is why nothing (get_addr?) has not converted the memory
addresses into VALUEs.

[Bug rtl-optimization/68955] [6 Regression] wrong code at -O3 on x86-64-linux-gnu in 32-bit mode

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68955

Jakub Jelinek  changed:

   What|Removed |Added

 Status|NEW |ASSIGNED
   Assignee|unassigned at gcc dot gnu.org  |jakub at gcc dot gnu.org

--- Comment #9 from Jakub Jelinek  ---
Created attachment 37362
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=37362=edit
gcc6-pr68955.patch

Fix, that passed bootstrap/regtest on x86_64-linux and i686-linux.  I'll
probably try another 2 bootstraps/regtests with some extra statistics
gathering, to find out how many times it triggeers a change in DSE behavior.

[Bug rtl-optimization/68955] [6 Regression] wrong code at -O3 on x86-64-linux-gnu in 32-bit mode

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68955

Richard Biener  changed:

   What|Removed |Added

   Priority|P3  |P1

[Bug rtl-optimization/68955] [6 Regression] wrong code at -O3 on x86-64-linux-gnu in 32-bit mode

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68955

Jakub Jelinek  changed:

   What|Removed |Added

 CC||jakub at gcc dot gnu.org

--- Comment #7 from Jakub Jelinek  ---
I believe this goes wrong during DSE2, at least before that pass we have always
6 sets of statements corresponding to
i[1][e][h] = i[h][k][e] >= l;
and then one corresponding to
i[e + 2][h + 3][e] = 6 & l;
and then one corresponding to
i[2][1][2] = a;
and all this 6 times.  But, after DSE2 all the
i[e + 2][h + 3][e] = 6 & l;
stores are removed, except the last one that is kept.
Sure, all the i[e + 2][h + 3][e] stores (in the same loop, i.e. same e and h)
are to the same address, and they don't alias with any other stores in the loop
(i[1][e][h] necessarily has smaller first index than any e + 2, and i[2][1][2]
has the second index smaller than any h + 3), but they might alias with the
i[h][k][e] loads.

[Bug rtl-optimization/68955] [6 Regression] wrong code at -O3 on x86-64-linux-gnu in 32-bit mode

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68955

Richard Biener  changed:

   What|Removed |Added

  Component|tree-optimization   |rtl-optimization

--- Comment #6 from Richard Biener  ---
Replacing the late DOM with FRE reproduces the issue as well so I suspect a
latent
issue elsewhere.  -fno-gcse also fixes this bug.